Supplying validation set manually

[sametsoekel]

Hi,

I'm aware of valid_split argument in tabnet_config() but it uses a proportion of train set as validation set. I have standalone validation sets, how can I use them during training goes on a separate data ?

Thanks in advance.

[cregouby]

Hello @sametsoekel There is not such a possibility for now in the package.

The easiest way to achieve it is to make a tabnet_fit() loop without split on your training-set, resuming the previous epoch weights via from_model= configuration option, and predict on your validation-set at each epoch to compute your val-loss. You will be able to see the end-to-end training loss with autoplot() as it is accumulated in the model.

The more performant way is to modify the training loop in https://github.com/mlverse/tabnet/blob/89a396ad5cdf449fa819d4ef7a6f92e30ef81da7/R/model.R#L358 for your need with your datasets

hope it helps,